Where is your POSEIDON now?

Ernesto Carrella

July 26, 2017

A minute of POSEIDON

  • POSEIDON is an agent-based model
    • Scalable
    • Modular
  • POSEIDON has a few use cases:
    • Scenario evalutation
    • Optimization
    • Reiforcement Learning

Work in progress

  • Calibrating and validating West-Coast DTS model
  • Develop the infrastructure for Indonesia
  • Data degradation for West-Coast fixed gear model

Calibration

  • We have two target data-sets:
    • Landings/quota attainments
    • Logbook data
  • Calibrate in two steps
    1. Fix agents statistically
    2. Find catchabilities that generate correct quotas
    3. Fix catchabilities but let agents act adaptively
    4. Find behavioural parameters that replicate logbook
  • How realistic can explore-exploit-imitate agents possibly be?

Calibration - Results

Calibration - Results 2

  • Looks realistic with simple adaptive agents
  • Adaptive agents are very risk-averse:
    • Calibrated exploration rate is 3.5%

Calibration - 15% exploration rate

Calibration - 3

  • It’s trivial to find more performing agents:
    1. Are we picking up an approximation error from the way we distribute fish compared to the real world?
    2. Are fishers just not that profit maximizing?

Indonesia

Fixed Gear Allocation